An experience-enabled text-to-CAD system recorded higher reported model-based scores than its no-experience comparison on CADFusion-Hard, a hard benchmark for shape descriptions in which some spatial constraints are left implicit. The paper targets executable CAD generation from natural-language descriptions by filling in those unstated constraints.
On CADFusion-Hard, the reported VLM score was 6.26 with Qwen3.5-27B and 6.60 with GPT-5. The VLM score is a model-based measure of text–shape consistency, gauging how closely a generated shape matches its description.
Against the same framework without experience, the memory-enabled variant was reported to have VLM performance 11.2% higher with Qwen3.5-27B and 7.5% higher with GPT-5. This was a computational ablation, and the paper reports no confidence intervals, standard deviations or significance tests.
From loose descriptions to executable code
The method is organized as a three-stage pipeline. It encodes retrieval queries and experience scopes with BGE bge-large-en-v1.5, then retrieves examples using cosine similarity.
The pipeline injects the top three retrieved experiences into generation and permits up to three rounds of code refinement. Its purpose is to help complete implicit spatial constraints while generating executable CAD code.
The experience bank was built from 1,500 randomly sampled instances in the CADFusion training set. During distillation, examples with initial scores lower than 7 were retained, and a regenerated experience was verified only when its score improved by more than 2 points.
How the hard cases were chosen
To construct each hard set, the researchers ran a baseline three independent times, selected 500 candidates with the lowest automatic scores, had them rated for difficulty on a 0-to-10 scale, and retained the 200 highest-difficulty samples from each benchmark. The resulting hard-set evaluations used 200 samples per benchmark.
The supplementary ablation discussion reports the full framework as the best-VLM version for both backbones. Removing construction guidance or implicit-constraint completion weakened performance.
The authors interpret the gap between the no-experience version and the full system as support for reusable experience. No uncertainty estimates were reported for those ablation comparisons.
Results changed with the test
On the full CADFusion test set, the pattern was less uniform. For Qwen3.5-27B, the reported VLM values were 7.56 without experience and 7.74 with it, while the authors described GPT-5 as slightly worse with experience. No uncertainty estimate was reported for this comparison.
Memory size also followed different patterns by backbone. Qwen3.5-27B reached its peak reported performance with 50 experience items and then saturated, whereas GPT-5 benefited from a larger memory.
On the unseen Text2CAD-Hard set, the supplied comparison used Qwen3.5-27B only. The experience-enabled system reported a VLM score of 6.54 versus 6.23 for Vanilla, while the reported Invalid Ratio values were 0.50 versus 52.00.
A metric check, and an unfinished memory
Separate validation examined whether the two-view VLM metric tracked human judgments. In 50 text–shape pairs rated by five annotators, its Spearman correlation was 0.84, compared with 0.71 for the three-view metric. The validation used a small volunteer-rated sample and reported no confidence interval or agreement statistic.
Experience management remains an open limitation. The paper did not study how stored experiences should be updated, deleted, merged or validated over time.
The work is an [arXiv preprint](https://arxiv.org/abs/2608.24760). Funding information is not reported. Its reproducibility section says source code will be publicly released and that the created artifacts are intended for research purposes.
Paper data and sources
Original title: ExpConCAD: Experience-Guided Text-to-CAD Generation from Shape Descriptions with Implicit Spatial Constraints
Authors: Jingyao Liu, Jinkang Tang, Chen Huang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text